样本数据预处理对基于BP神经网络的GPS高程拟合的影响

INFLUENCE OF SAMPLE DATA PREPROCESSING ON BP NEURAL NETWORK-BASED GPS ELEVATION FITTING

分析在基于BP神经网络的GPS高程拟合建模中样本数据预处理的必要性,并列举了归一化、中心化、标准化3种数据预处理方法,然后结合实例,在神经网络建模中增加一个数据预处理层,分别用3种训练方法对基于不同数据预处理的模型进行训练建模,将计算结果进行对比分析,并与二次曲面模型结果进行比较,得出不同数据预处理方法对基于神经网络的GPS高程拟合建模精度的影响不同,且神经网络方法比二次曲面方法的拟合精度更高。

The necessity of sample data preprocessing for BP neutral network-based GPS elevation fitting is analyzed,and three preprocessing methods are given,which are normalization,centralization and standardization.In order to obtain an optimal model,a data preprocessing layer is added in the network topology and then the three different methods are applied in training the neutral networks.Through the comparative analysis of the results derived with different preprocessing methods and with quadratic surface model,it is found that the influence of three data preprocessing methods on BP neutral network-based GPS elevation fitting is different,and the accuracy with BP neutral network method is higher than that with quadratic surface model.